Alleviate of rendering appearing noisy due to sh3 precision limitations

[heimeii]

#142 (comment)

encode sh3 to sint6 min precision is 1/31 (0.03225)

if use uin8 replace, min precision is 1/127 (0.00787)

Based on my current observations, the value of sh3 frequently tends to be small, which is highly likely the cause of this phenomenon.

in sh3 buffer, item total size is 128 and used 126, 2bit is free

so, we can real calculate scale to alleviate precision.

code look like (only test code):

before:

after:

Although there has been some improvement, it still falls short of supersplat level of cleanliness

[heimeii]

after opt parse perf in wasm is 1.2s, spark now is 2.3

[mrxz]

Could you try with #146? It increases the SH3 coefficients from sint6 to sint8 which should allow for higher precision.

[heimeii]

Could you try with #146? It increases the SH3 coefficients from sint6 to sint8 which should allow for higher precision.

Thank you for the implementation suggestions, though they might not be suitable for my scenario. I suspect they could cause performance degradation on low-end devices, as GPU texture fetch caching is relatively unfriendly to linear memory access patterns. I'll need to wait until Monday to test this empirically.